thesreedath/slm125m-live-base
016
thesreedath/slm125m-live-base
A 125M-parameter Llama-architecture small language model pretrained from scratch on a legal/financial corpus (US case law + SEC filings + a fineweb-edu web slice), built with the slm125mLIVE pipeline.
- Params: ~125.8M (12L / 768d / 12h, seq 1024, tied embeddings)
- Tokenizer: 16,384 byte-level BPE trained on the corpus
- Training tokens: ~2.50B (~1 epoch over ~2.19B unique)
- Final val loss: 2.3132250440120696 | val ppl: 10.106967565875875
- Mix: ~40% case-law / ~40% SEC / ~20% web
